Скачать или смотреть CSCI 1109 - M40 - Baselines & leakage traps

In this module, we stop cheering for big accuracy numbers and start asking the harder question: “compared to what, and can we trust it?” You’ll learn how to build simple but powerful baseline models, see how easy it is to accidentally leak information from the future into your training process, and treat model evaluation as a mini scientific experiment rather than a screenshot of one lucky split.

Build and interpret constant and majority-class baselines for regression and classification, and use them as a realistic “floor” any serious model must beat.
Compare a baseline to a k-NN or logistic regression model using confusion matrices and metrics, and reason about when extra complexity is actually worth it.
Spot common forms of data leakage—train/test contamination, temporal leakage, and target proxies—and see how they can create “too good to be true” results.
Refactor small sklearn workflows with proper train/validation/test handling, pipelines, and (lightweight) cross-validation so preprocessing happens only on training data.
Connect these practices to real scientific failures and retractions, and articulate simple, practical checks that make your own work more reproducible and trustworthy.
By the end of the module, you should be able to explain why every predictive model needs a baseline, recognize red flags that suggest data leakage, and read model performance numbers with a more skeptical, science-minded eye. In other words, you’ll be better equipped to tell the difference between a genuinely useful model and a very fancy cheat.

Course module page: https://web.cs.dal.ca/~rudzicz/Teaching/CS...